https://www.newyorker.com/magazine/2026/02/16/what-is-claude-anthropic-doesnt-know-either
Of late, in most conversations I am part of on AI for impact, everyone talks as if we already know what these systems are. I am not sure we do, and if someone has figured it out, they have not been honest about it. This New Yorker piece, which I read a while back, is a fascinating read and forces you to revisit a few assumptions.
- The people who built Claude say openly that they do not fully understand it. They are studying it the way one would study a brain, by looking at its neurons and running it through psychology experiments. That tells you something about where the field actually is.
- I had been repeating the line that these models just predict the next word. The article shows that this is technically true but practically misleading. The model holds concepts and plans ahead.
- And then there is the story of Claude running a vending machine inside Anthropic’s office. It hallucinated a phone call, invented an office address, and got fooled by made-up discount codes. If you are thinking about putting an AI agent in charge of anything in a DFI or a social enterprise, read that section twice.
This is a long article but surely worth your time (don’t put this in an AI program and ask for a summary.. the long-form reading has its own advantages).
